PaperlayerView exact revision on arXiv

Source-linked AI summary

Structure-sensitive Multi-scale Deep Neural Network for Low-Dose CT Denoising

Chenyu You, Qingsong Yang, Hongming Shan, Lars Gjesteby, Guang Li, Shenghong Ju, Zhuiyang Zhang, Zhen Zhao, Yi Zhang, Wenxiang Cong, Ge Wang

arXiv:1805.00587v3cs.CVcs.AI

TL;DR

LDCT lowers radiation exposure but introduces noise and artifacts that can impair image quality. The paper proposes SMGAN, a 3D structure-sensitive generative adversarial network using volumetric information and combined losses. Experiments report preservation of structural and texture information, noise and artifact suppression, and better detailed visual results than competing methods.

  • Problem

    Radiation-dose reduction increases LDCT noise and artifacts, creating a challenge for maintaining diagnostic image quality.

  • Method

    SMGAN uses 3D volumetric processing with a structurally sensitive loss and adversarial training to denoise LDCT images.

  • Results

    Experiments show that SMGAN preserves structural and texture information, suppresses noise and artifacts, and outperforms competing methods in radiologists' qualitative assessments.

  • Takeaways & Limitations

    The method produces higher-level image quality for clinical usage while retaining detailed information for diagnosis.

Abstract

from arXiv · show

Computed tomography (CT) is a popular medical imaging modality in clinical applications. At the same time, the x-ray radiation dose associated with CT scans raises public concerns due to its potential risks to the patients. Over the past years, major efforts have been dedicated to the development of Low-Dose CT (LDCT) methods. However, the radiation dose reduction compromises the signal-to-noise ratio (SNR), leading to strong noise and artifacts that down-grade CT image quality. In this paper, we propose a novel 3D noise reduction method, called Structure-sensitive Multi-scale Generative Adversarial Net (SMGAN), to improve the LDCT image quality. Specifically, we incorporate three-dimensional (3D) volumetric information to improve the image quality. Also, different loss functions for training denoising models are investigated. Experiments show that the proposed method can effectively preserve structural and texture information from normal-dose CT (NDCT) images, and significantly suppress noise and artifacts. Qualitative visual assessments by three experienced radiologists demonstrate that the proposed method retrieves more detailed information, and outperforms competing methods.

I. INTRODUCTION

LDCT reduces radiation exposure but increases noise and can compromise diagnostic information. SMGAN addresses this challenge with a 3D adversarial denoising framework that combines structural, perceptual, and pixel-wise objectives.

  • Reducing radiation dose increases image noise and can severely compromise diagnostic information in LDCT.
  • Sinogram filtering methods are computationally efficient but may lose structural information and spatial resolution.
  • Mean-squared-error-based denoising can reduce noise and artifacts but may produce blurry results.
  • SMGAN combines adversarial, perceptually favorable structural, and pixel-wise L1 losses to capture subtle structural features while maintaining visual sensitivity.
  • The proposed network maps volumetric LDCT images toward the NDCT feature space while a discriminator distinguishes synthetic from real NDCT images.

B. 3D Spatial Information

The method uses 3D spatial context and a structurally sensitive loss within an SMGAN architecture. Its generator processes volumetric LDCT patches, while the discriminator evaluates generated images against NDCT references.

  • B. 3D Spatial Information: Adjacent CT slices exhibit strong spatial correlation that 3D networks can use to preserve more information than 2D models.
  • B. 3D Spatial Information: SMGAN-3D uses a 3D ConvNet generator and 3D Structurally-Sensitive loss on volumetric LDCT patches.
  • C. Network Structure: The generator has eight 3D convolutional layers and avoids pooling to reduce loss of subtle textural and structural information.
  • C. Network Structure: The 3D SSL function measures patch-wise discrepancy between generated and NDCT volumes in the spatial domain and back-propagates it to update the network.
  • C. Network Structure: The discriminator uses six convolutional layers followed by two fully connected layers to distinguish generated images from real NDCT images.

3) Discriminator:

The experiments compare loss functions for LDCT denoising, showing different trade-offs between noise suppression, blurring, and artifacts. L1 loss reduces blurring relative to L2, but does not eliminate blocky artifacts.

  • Loss-function comparison: The study evaluates different loss functions to justify using a hybrid loss for optimal diagnostic quality.The compared components include L1, structural, and Wasserstein-based losses.
  • Loss-function comparison: L2 loss suppresses background noise efficiently but can produce unnatural, blurry results through regression to the mean.It also assumes white Gaussian noise independent of local image features.
  • Loss-function comparison: L1 loss avoids over-penalizing large differences and can alleviate drawbacks associated with L2 loss.Its main limitation is nondifferentiability, unlike L2 loss.
  • Loss-function comparison: Compared with L2 loss, L1 loss suppresses blurring but does not help reduce blocky artifacts.This comparison is reported from the visual results in Figs. 3–6.

3) Adversarial Loss:

The adversarial-loss formulation uses Wasserstein distance with gradient penalty, while structural similarity measures capture image relationships across scales. These components provide differentiable structural guidance for denoising optimization.

  • Wasserstein adversarial loss: The Wasserstein adversarial loss combines two distribution-distance terms with a gradient-penalty term.The gradient penalty regularizes the discriminator, and the generated image is denoted by z.
  • Structural similarity: SSIM measures structural and perceptual similarity using means, standard deviations, and cross-covariance between generated and NDCT images.These metrics are motivated by visual pattern recognition and exploit feature correlations in medical images.
  • Structural similarity: MS-SSIM extends structural similarity to multiscale analysis across local image-content levels.SSIM is a special case of MS-SSIM.
  • Structural loss: The structural loss is differentiable, allowing its value to be back-propagated to update network weights.This makes the structural discrepancy usable directly in network training.

5) Objective Function:

The objective combines structural-sensitive and L1 losses with adversarial learning to balance structure preservation, noise suppression, and realistic image features. The resulting error is back-propagated to optimize the network.

  • Composite objective: L1 loss supports noise suppression and SNR improvement, whereas structural loss discourages blurring and preserves high contrast resolution.The contrast motivates combining the two losses.
  • Composite objective: The structural-sensitive loss combines structural loss and L1 loss using τ to balance structure preservation against noise suppression.The weighting factor τ controls the contribution of the two terms.
  • Composite objective: Adversarial loss is added because the two reconstruction losses may miss diagnostic features, helping retain textural and structural features.The overall SMGAN objective therefore includes a weighted adversarial component.
  • Composite objective: The network compares its output volume with the target volume and back-propagates the resulting error for optimization.β specifies the adversarial-loss weight in the overall objective.

III. EXPERIMENTS AND RESULTS

Experiments use Mayo Clinic low-dose CT data, overlapping 3D patches, and comparisons against eight denoising methods using visual and quantitative evaluation. Results examine convergence and image quality across 2D, 3D, mean-based, structural-loss, and adversarial approaches.

  • Experimental datasets and setup: The Mayo dataset contains 2,378 NDCT and quarter-dose LDCT images from 10 anonymous patients.The reconstruction interval is 0.8 mm and slice thickness is 1.0 mm.
  • Experimental datasets and setup: Overlapping 3D patches are used to capture patch-wise spatial interconnections and enlarge the training patch dataset.The model processes 80 × 80 × 11 patches extracted with an overlapping sliding window.
  • Experimental comparisons: SMGAN-2D and SMGAN-3D are qualitatively compared with eight methods, including CNN-L2, CNN-L1, SL-net, MSL-net, WGAN, BM3D, RED-CNN, and WGAN-VGG.The comparison includes classical, pixel-wise, perceptual-loss, structural-loss, and adversarial approaches.
  • Experimental comparisons: Image quality is quantitatively evaluated with PSNR, SSIM, and RMSE.These metrics complement the qualitative comparisons.
  • Network convergence: L1-net and L2-net converge fastest, with both curves converging around the 6th epoch in the L1-loss comparison.Their curves initially decrease and then smoothly converge.
  • Network convergence: SMGAN-2D has the mildest oscillation among the compared Wasserstein models and reaches a stable state after the 13th epoch.SMGAN-3D oscillates over a larger range because it incorporates 3D structural information, while remaining close to the x-axis.

D. Denoising Performance

Across abdominal CT examples, denoising methods suppress noise to varying degrees, but SMGAN methods better preserve structural details and edges than competing approaches.

  • All evaluated denoising models suppress noise relative to LDCT, which is noisy and lacks structural features needed for clinical diagnosis.
  • Mean-based methods reduce noise effectively but can blur crucial structural information and impair image contents.
  • SL-based methods preserve texture and contrast resolution better than mean-based methods, but may lose overall intensity and retain structural artifacts.
  • SMGAN-based methods provide better edge preservation than competing methods while maintaining noise suppression and artifact removal.
  • SMGAN-3D demonstrates benefits from combining two loss functions with adversarial training for structural and texture preservation.

2) Comparison with WGAN-based denoising methods:

Compared with WGAN-based alternatives, SMGAN-3D combines adversarial learning with a structurally sensitive objective and 3D information to improve anatomical detail preservation.

  • Comparison with WGAN-based denoising methods:: The comparison evaluates SMGAN methods against WGAN and WGAN-VGG, focusing on clinical image quality and structural features.
  • Comparison with WGAN-based denoising methods:: WGAN can capture structural information, but using it alone may produce stronger noise because it does not model local voxel intensity and structural correlations.
  • Comparison with WGAN-based denoising methods:: WGAN-VGG provides more perceptual detail but may severely distort original structural information, possibly because VGG is pretrained on natural images.
  • Comparison with WGAN-based denoising methods:: SMGAN-3D more clearly visualizes metastasis and better preserves the portal vein than WGAN and WGAN-VGG.
  • Comparison with WGAN-based denoising methods:: SMGAN-based methods achieve better anatomical feature preservation and visual quality than other state-of-the-art methods in the illustrated comparisons.
  • Comparison with WGAN-based denoising methods:: The proposed objective function is reported as essential for capturing more accurate anatomical details.

3) Comparison with Image space denoising:

The paper compares SMGAN-3D with image-space denoising and evaluates quantitative, statistical, and reader-assessed image quality. SMGAN-3D is reported to preserve anatomical details while suppressing noise and artifacts.

  • Comparison with Image space denoising:: BM3D blurs a low-contrast lesion and smooths specific features, whereas SMGAN-3D produces sharper features in the corresponding ROIs.
  • Comparison with Image space denoising:: SMGAN-3D produces better subtle details and more similar statistical noise properties to NDCT than SMGAN-2D.
  • Comparison with Image space denoising:: SMGAN-3D incorporates 3D structural information, while slice-wise SMGAN-2D may lose spatial correlation between adjacent slices.
  • Comparison with Image space denoising:: SMGAN-3D provides improved anatomical feature preservation over other state-of-the-art methods in the illustrated comparisons.
  • Comparison with Image space denoising:: The evaluation combines PSNR, SSIM, RMSE, statistical image properties, and a blind visual assessment by three radiologists.
  • Comparison with Image space denoising:: RED-CNN achieves the highest PSNR and RMSE, while SL-net and MSL-net achieve the highest SSIM scores.
  • Comparison with Image space denoising:: The authors caution that standard metrics may not sufficiently indicate structural detail preservation or diagnostic performance because they are affected by blurring and artifacts.
  • Comparison with Image space denoising:: SMGAN-2D outperforms WGAN-VGG on the three selected metrics, and SMGAN-3D achieves higher PSNR and SSIM than SMGAN-2D.

2) Statistical analysis:

Quantitative and visual assessments compare denoising models against NDCT, showing that SMGAN-3D best matches NDCT variability and performs strongly across key clinical image-quality criteria.

  • Statistical analysis:: Lower percent errors in ROI mean and standard-deviation values indicate more robust denoising models.The analysis compares each model’s mean CT number and SD against NDCT.
  • Statistical analysis:: SMGAN-3D achieves the best match to NDCT standard deviations among the evaluated denoising methods.WGAN, WGAN-VGG, SL-net, and SMGAN-2D reduce noise but deviate from NDCT mean values and compromise content information.
  • Visual assessments:: Three radiologists evaluate ten image groups for sharpness, noise suppression, diagnostic acceptability, and contrast retention using five-point scores.Each group contains LDCT, NDCT, and outputs from different denoising methods.
  • Visual assessments:: SMGAN-3D scores best for sharpness, diagnostic acceptability, and contrast retention, while RED-CNN scores highest for noise suppression.SMGAN-3D also achieves a superior overall image-quality score relative to SMGAN-2D.
  • Visual assessments:: Visual assessment indicates that SMGAN-3D reduces noise and artifacts while preserving subtle image structures, edges, and HVS requirements.The results are illustrated in ROI comparisons and assessment figures.

F. Computational Cost

The study examines computational cost and loss-function trade-offs in LDCT denoising. Training is time-consuming, but trained models denoise rapidly and the proposed hybrid loss integrates complementary strengths.

  • Computational Cost: DL-based denoising requires time-consuming training but can rapidly process reconstructed LDCT images after training.This supports lower execution overhead during clinical workflow than iterative reconstruction.
  • Computational Cost: 15 hours for SMGAN-2D and approximately 26 hours for SMGAN-3D are required for training to converge.WGAN-VGG takes about 18 hours with the same number of layers.
  • Computational Cost: 0.534s for SMGAN-2D and 4.864s for SMGAN-3D are reported for validation on an NVIDIA Titan GPU.The cited comparison reports 45.1s per iteration for STF-based TV minimization in OS-SART on the same platform.
  • Loss-function trade-offs: Mean-based losses provide strong quantitative scores and noise reduction, whereas feature-based methods preserve more structural detail but risk content distortion.The discussion attributes the risk to perceptual loss computed from a network trained on natural images.
  • Loss-function trade-offs: The proposed hybrid loss combines complementary loss-function characteristics to optimize LDCT image quality.Adversarial learning can prevent smoothing and capture structural characteristics, but may also lose diagnostic information.
  • Limitations: Remaining issues include blurry feature edges, imperfect NDCT–LDCT structural matching, and reader-study variability.The authors plan to improve network modeling and quantify inter- and intra-operator variability with more radiologists.
  • Conclusion: The authors report higher-level image quality for clinical usage compared with existing denoising networks.The conclusion is based on quantitative and qualitative image-quality comparisons.
  • Future work: Future work includes extending the model to other modalities, incorporating advanced denoising models, and improving robustness across scanners.These directions define the reported scope for continued development.

APPENDIX A DIFFERENT TRAINING SETS FOR SMGAN-3D TRAINING

The appendix tests SMGAN-3D generalizability across four Mayo-dataset training sets, each containing 5,000 volumetric patches, with results reported in a figure and table.

  • APPENDIX A DIFFERENT TRAINING SETS FOR SMGAN-3D TRAINING: The Mayo dataset is randomly split into four training sets to validate SMGAN-3D generalizability.The results are presented in Fig. 10 and Table IV.
  • APPENDIX A DIFFERENT TRAINING SETS FOR SMGAN-3D TRAINING: Each training set contains 5,000 image patches sized 80 × 80 × 11 pixels.The patches are used for SMGAN-3D training-set comparisons.
  • APPENDIX A DIFFERENT TRAINING SETS FOR SMGAN-3D TRAINING: Table IV summarizes quantitative results associated with the different SMGAN-3D training sets.Fig. 10 displays the corresponding results.
  • Notation: The notation summary defines NDCT, LDCT, SSL, SSIM, MS-SSIM, and the evaluated denoising models.SSL integrates structural loss and L1 loss as defined in Eq. 10.
Loading 1805.00587v3…